Can Model Context Protocol (MCP) Transform AI Agent Tooling?

What is MCP?

MCP (Model Context Protocol) is an open protocol that standardizes how AI models call external tools, retrieve data, and interact with services. Inspired by the Language Server Protocol (LSP), MCP extends the concept with an agent‑centric execution model, allowing autonomous AI workflows to decide which tools to use, in what order, and how to chain calls.

Current mainstream use cases

When an MCP server is available, each MCP client can become an "everything app". For example, the code editor Cursor can act as a full MCP client: installing a Slack MCP server turns it into a Slack client, a Resend MCP server enables email sending, and a Replicate MCP server generates images. Multiple MCP servers can be combined to create composite tasks, such as generating a front‑end UI and a hero image in one workflow.

Most existing use cases fall into two categories:

Developer‑focused localized workflows

New consumer experiences built on LLM clients

1. Developer‑focused workflows

Developers no longer need to leave their IDE to check a database, manage caches, or view logs. A Postgres MCP server lets them run read‑only SQL directly in the editor; an Upstash MCP server handles cache indexes; Browsertools MCP provides real‑time console access for debugging.

Beyond basic tool calls, MCP servers can automatically generate context‑aware code snippets by scraping webpages or generating servers from documentation, reducing boilerplate and letting developers focus on higher‑level logic.

2. New consumer experiences

Non‑technical users can access AI tools through friendly MCP clients such as Claude Desktop. Future clients may target customer support, marketing copy, design, or 3D modeling, leveraging AI’s strengths in pattern recognition and creative generation.

Client design determines functional boundaries: a chat app won’t expose a vector canvas, and a design tool won’t execute code remotely. This leaves ample room for innovative UI/UX patterns.

Highlight’s “@command” feature illustrates how an MCP client can push generated content to any downstream application.

Blender MCP servers enable novices to describe 3D models in natural language, hinting at a growing text‑to‑3D workflow ecosystem that will eventually include Unity and Unreal Engine.

Future possibilities

Developer‑centric companies may shift competitive advantage from API design to curated tool collections that agents can automatically discover and use.

Every application could become an MCP client, and every API an MCP server, leading to dynamic pricing based on speed, cost, and relevance.

Machine‑readable documentation (e.g., llms.txt) will become essential for exposing tool capabilities to agents.

APIs will evolve from endpoints to orchestrated workflows, where a single high‑level call like draft_email_and_send() internally invokes multiple services.

Hosting models will adapt to support multi‑step, resumable, and retryable executions required by AI agents.

Key challenges and open problems

1. Multi‑tenancy and hosting

Supporting many users on a single MCP server requires robust, SaaS‑grade tooling for deployment, scaling, and separation of data and control planes.

2. Authentication

MCP currently lacks a standardized identity mechanism. Implementations rely on ad‑hoc OAuth or API‑token schemes, which are insufficient for large‑scale remote deployments.

3. Authorization

Fine‑grained permission models are missing; most implementations use a binary session‑level allow/deny approach, often based on OAuth 2.1 flows.

4. Gateway layer

As MCP ecosystems grow, a dedicated gateway could centralize auth, routing, traffic management, and load balancing, similar to traditional API gateways.

5. Discoverability and availability

Finding and configuring MCP servers remains cumbersome. Upcoming registration and discovery protocols (e.g., announced by Anthropic) could streamline this process.

6. Execution environment

MCP does not define a native "workflow" concept, forcing each client to implement its own resumability and retry logic. Integrating existing workflow engines (e.g., Inngest) could alleviate this burden.

7. Standardized client experience

There is no unified UI/UX for tool selection—some clients use slash commands, others natural‑language triggers. A common discovery, ranking, and execution interface would improve consistency.

8. Debugging

Debug information is often missing across different MCP clients, making cross‑client troubleshooting difficult. Better observability tools are needed, especially as remote MCP servers become prevalent.

Impact on the AI tool ecosystem

MCP is reshaping how AI agents interact with tools, similar to how APIs standardized software communication in the early web era. If widely adopted, MCP could drive new business models, encourage machine‑readable documentation, and push hosting providers to support multi‑step, stateful AI workflows.

Source: a16z.com/a-deep-dive-into-mcp-and-the-future-of-ai-tooling/